CoChat

Detecting Deepfake Voice Using Explainable Deep Learning Techniques

Open access
Lim, Suk-Young, Chae, Dong-Kyu, Lee, Sang-Chul · 2022 · Applied Sciences

The DOI matches its Crossref record; 21 of the 21 references it cites were found in a registry or an index.

Abstract

Fake media, generated by methods such as deepfakes, have become indistinguishable from real media, but their detection has not improved at the same pace. Furthermore, the absence of interpretability on deepfake detection models makes their reliability questionable. In this paper, we present a human perception level of interpretability for deepfake audio detection. Based on their characteristics, we implement several explainable artificial intelligence (XAI) methods used for image classification on an audio-related task. In addition, by examining the human cognitive process of XAI on image classification, we suggest the use of a corresponding data format for providing interpretability. Using this novel concept, a fresh interpretation using attribution scores can be provided.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.3390/app12083926

Cites

21 of 21 references verified0 citation errors · 0 not checkable

Statistics

Figures

Figure 1 Deep Taylor, integrated gradients, and LRP on image classification through VGG16. ( a ) Input. ( b ) Deep Taylor. ( c ) IG. ( d ) LRP.
Figure 1 Deep Taylor, integrated gradients, and LRP on image classification through VGG16. ( a ) Input. ( b ) Deep Taylor. ( c ) IG. ( d ) LRP.CC-BY · From the PDF
Figure 2 Illustration of CNN-LSTM Model Structure.
Figure 2 Illustration of CNN-LSTM Model Structure.CC-BY · From the PDF
Figure 3 ASVspoof attribution score with the CNN model. ( a ) Deep Taylor deepfake voice. ( b ) Integrated gradients deepfake voice. ( c ) LRP deepfake voice. ( d ) Deep Taylor real voice. ( e ) Integrated gradients real voice. ( f ) LRP real voice.
Figure 3 ASVspoof attribution score with the CNN model. ( a ) Deep Taylor deepfake voice. ( b ) Integrated gradients deepfake voice. ( c ) LRP deepfake voice. ( d ) Deep Taylor real voice. ( e ) Integrated gradients real voice. ( f ) LRP real voice.CC-BY · From the PDF
Figure 4 ASVspoof attribution score with CNN-LSTM model. ( a ) Deep Taylor Deepfake voice. ( b ) Integrated gradients deepfake voice. ( c ) LRP deepfake voice. ( d ) Deep Taylor real voice. ( e ) Integrated gradients real voice. ( f ) LRP real voice.
Figure 4 ASVspoof attribution score with CNN-LSTM model. ( a ) Deep Taylor Deepfake voice. ( b ) Integrated gradients deepfake voice. ( c ) LRP deepfake voice. ( d ) Deep Taylor real voice. ( e ) Integrated gradients real voice. ( f ) LRP real voice.CC-BY · From the PDF
Figure 5 LJSpeech attribution score with CNN model. ( a ) Deep Taylor deepfake voice. ( b ) Integrated gradients deepfake voice. ( c ) LRP deepfake voice. ( d ) Deep Taylor real voice. ( e ) Integrated gradients real voice. ( f ) LRP real voice.
Figure 5 LJSpeech attribution score with CNN model. ( a ) Deep Taylor deepfake voice. ( b ) Integrated gradients deepfake voice. ( c ) LRP deepfake voice. ( d ) Deep Taylor real voice. ( e ) Integrated gradients real voice. ( f ) LRP real voice.CC-BY · From the PDF
Figure 6 LJSpeech attribution score with CNN-LSTM model. ( a ) Deep Taylor deepfake voice. ( b ) Integrated gradients deepfake voice. ( c ) LRP deepfake voice. ( d ) Deep Taylor real voice. ( e ) Integrated gradients real voice. ( f ) LRP real voice.
Figure 6 LJSpeech attribution score with CNN-LSTM model. ( a ) Deep Taylor deepfake voice. ( b ) Integrated gradients deepfake voice. ( c ) LRP deepfake voice. ( d ) Deep Taylor real voice. ( e ) Integrated gradients real voice. ( f ) LRP real voice.CC-BY · From the PDF
Figure 7 LJSpeech attribution score with CNN-LSTM-perm model. ( a ) Deep Taylor deepfake voice. ( b ) Integrated gradients deepfake voice. ( c ) LRP deepfake voice. ( d ) Deep Taylor real voice. ( e ) Integrated gradients real voice. ( f ) LRP real voice.
Figure 7 LJSpeech attribution score with CNN-LSTM-perm model. ( a ) Deep Taylor deepfake voice. ( b ) Integrated gradients deepfake voice. ( c ) LRP deepfake voice. ( d ) Deep Taylor real voice. ( e ) Integrated gradients real voice. ( f ) LRP real voice.CC-BY · From the PDF
Figure 8 Spectrogram of LJSpeech 'The wildest and most cut-throat looking of the lot, which proves that he could be grateful for kindness, and was not all bad.' ( a ) Deepfake voice. ( b ) Real voice.
Figure 8 Spectrogram of LJSpeech 'The wildest and most cut-throat looking of the lot, which proves that he could be grateful for kindness, and was not all bad.' ( a ) Deepfake voice. ( b ) Real voice.CC-BY · From the PDF
Figure 9 Top 10% amplitude of reconstructed Deep Taylor-based attribution score from LJSpeech. ( a ) Deepfake voice. ( b ) Real voice.
Figure 9 Top 10% amplitude of reconstructed Deep Taylor-based attribution score from LJSpeech. ( a ) Deepfake voice. ( b ) Real voice.CC-BY · From the PDF

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Research works printed but not linked: 2

  1. Bracewell, R.N., and Bracewell, R.N. (1986). The Fourier Transform and Its Applications, McGraw-Hill.
  2. Becker, S., Ackermann, M., Lapuschkin, S., Müller, K.R., and Samek, W. (2018). Interpreting and explaining deep neural networks for classification of audio signals. arXiv.

Cited by

See all 83

Metadata

Authors
Lim, S.-Y., Chae, D.-K., & Lee, S.-C.
Affiliations
  • Hanyang University (KR)
  • Division of Nanotechnology, Daegu Gyeongbuk Institute of Science & Technology (DGIST), Deagu 42988, Korea
Venue
Applied Sciences
Publisher
MDPI AG
Published
2022-04-13
Volume
12
Issue
8
Pages
3926
ISSN
2076-3417 (ISSN-L)
OpenAlex
W4223915588
Cited by
82
References
17
Citations by year
2026: 17 · 2025: 31 · 2024: 23 · 2023: 10 · 2022: 1
FWCI
5.35
Type
journal article
Language
en
Open access
Locations
  • Applied Sciences (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • Korea government(MSIT) (No.2020-0- 378 01373)
  • Korean government (MSIT) (No. NRF-2021M3E5D2A01021156)
  • Ministry of Science and ICT of KOREA (21-IT-10-03)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
  • Physical Sciences › Computer Science › Artificial Intelligence › Speech Recognition and Synthesis
Keywords
deepfake detection, explainable AI, audio forensics, interpretability, fake media, audio classification

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.